The owner's framing, and why the old answer was too narrow

"i said — there's no shame in a certain garbage collection. you concluded it wasn't needed, since — procedure exit — clean up temps. but it's both sides of the same medal. sometimes it would be easier to just alloc memory, tag it, use it, have it collect after we proven it's no longer in use. this (thread-safe) pchar conversion would be a pretty typical use case." — owner, 2026-09-14

He is right that they are one medal. IRParkManagedStr (compiler/ir.inc:4528) allocates, tags, uses and collects: the tag is a refcount, and the proof is the enclosing scope's exit. That is his scheme with a per-object tag instead of a per-region one.

What that choice costs, visible in the code: ~130 lines across IRParkManagedStr / IRParkManagedDyn, seven managed-string-temp seams, and a whole twin of the mechanism — the comment at ir.inc:4595 says why the string park could not see dyn arrays, "a dyn-array-typed node already reads tyPointer". Most of that complexity answers one question, who owns this temp: IRParkManagedStr "returns val UNCHANGED when the value is ALREADY OWNED", and getting that predicate wrong is a double free in one direction and a leak in the other. A region never asks the question. The region owns it.

The case refcounting cannot serve at all

A refcount needs a handle. Every conversion that must build bytes has none: append a NUL, transcode to UTF-16, render a format. There is no object to tag, so the only answer in the tree is a file-scope static — which is what lib/pcl/gtk3.pas's PC() was until 2026-09-14, four 1024-byte slots, and it had both failure modes a static has: it overran, and it collided with itself inside one call. See [[bug-b-gtk3-pc-writes-past-its-buffer-on-a-long-string]].

What was MEASURED, and it shrinks the case rather than making it

Three probes, 2026-09-14, all NEGATIVE for a hole:

probe	result
managed string, lengths 0..4096, 3 construction routes	NUL always present (`builtinheap.pas:2386` reserves it)
`PChar(f)` for `f: string[16] := 'abc'`	`strlen` = 3, NUL ok
`PChar(a)` for `a: string[16]` filled to exactly 16	`strlen` = 16, NUL ok

So the RTL's own string seams do not currently need this. The population that bit is hand-written bindings, where somebody reached for a static because the language offered nothing else. That is the honest size of it today, and it is the number to re-measure before anyone builds anything.

The open question is the RECLAIM POINT

Not the allocator — a per-thread bump pointer is a morning's work and is thread-safe by construction, which is the owner's parenthesis and the part that matters most after the CLONE_SETTLS week.

Statement end is the tight bound and matches what FPC guarantees for PChar of a temporary. Cheap to prove: the compiler knows where a statement ends.
Scope exit is the conservative fallback and is what the park already does.

Either is only sound if nothing captured a pointer into the region — the same escape question every region scheme has, and the reason this is a ticket and not a patch. Do not start with the allocator. Start by deciding which of the two bounds is provable with what the compiler already knows, because that decides whether the thing is worth having.